OpenAI GPT-6 Cyber: 81 Firefox Flaws, 948 Linux Exploits
AI-assisted security research can expose weaknesses before attackers exploit them — but the same tools can also produce working attack code. OpenAI says its new
AI-assisted security research can expose weaknesses before attackers exploit them — but the same tools can also produce working attack code. OpenAI says its new restricted-access GPT-6 Cyber model found 81 previously unknown high-severity Firefox flaws and built 948 working exploits for Linux kernel vulnerabilities.
The claims point to a model designed to find software bugs, test whether they can be exploited and suggest fixes. Mozilla’s severity ratings provide external confirmation for the Firefox findings, while the broader figures remain largely based on OpenAI’s account.
What the model found
OpenAI researchers used GPT-6 Cyber to generate reproducible tests for 81 previously unknown Firefox flaws. Mozilla classified them as high severity, and OpenAI sent the company its findings and proposed patches. Mozilla has fixed many of the vulnerabilities; another account of the work says the 81 have been patched.
The model also indicated that 30 previously known Firefox issues were more serious than earlier assessments suggested. OpenAI says the new findings were previously unknown, but the 81-count itself is the company’s claim; Mozilla’s severity assessment is the third-party validation described in the available reporting.
The Linux work tested a different kind of risk. In an isolated environment, researchers reproduced 1,724 kernel vulnerabilities and developed working exploits for 948. The exploits demonstrated how someone starting with low-privilege access could gain root privileges, which provide administrative control.
Capability and safeguards
OpenAI reported 304 critical Linux findings to maintainers, along with candidate patches. The research campaign cost about $1.2 million in API credits, a measure of the resources involved in testing at this scale.
On OpenAI’s ExploitGym benchmark, which challenges models to exploit a specified flaw on a remote target, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. In a separate evaluation by security firm Irregular, it completed all 11 multi-stage attack scenarios at least once.
OpenAI classifies GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, a higher rating than the High classification assigned to GPT-5.6 Cyber. Access to the model is restricted.
The company also says it trained the model to follow operational boundaries, including with a honeypot inspired by a previous security incident. In tests without special safety instructions, it stayed within its assigned limits in 99.93% of cases, up from 94.49% for its predecessor; with instructions enabled, OpenAI reported a 100% rate.
For software makers, the findings show how AI could help teams reproduce bugs and prepare patches, while the working exploits underline the potential for misuse. OpenAI’s Linux researchers reported 304 critical findings to maintainers.
